Enhancing LiDAR place recognition in challenging environments using Retrieval-Trigger-Reranking paradigm

LiDAR place recognition (LPR) plays a critical role in simultaneous localization and mapping (SLAM) and autonomous driving systems. However, the common separation between offline database construction and online localization introduces challenges such as rotational variance, sensor discrepancies, and long-term environmental changes. Existing methods relying on fixed-length global descriptors often struggle in such scenarios due to their inherently limited capacity to encode comprehensive environmental information. To address these challenges, we propose RTR-Net, a novel framework based on Retrieval-Trigger-Reranking paradigm, to enhance LPR performance in challenging environments. The framework operates in three phases: (1) Retrieval, where a lightweight backbone generates global descriptors and local regional features for initial candidate selection; (2) Trigger, a training-free module that assesses spatial consistency between query and candidates to activate reranking only when necessary; and (3) Reranking, which refines rankings by fusing local features, global descriptors, and spatial consistency scores via spatial and channel attention mechanisms. Additionally, a regional sampling method is proposed to mitigate field-of-view (FoV) discrepancies across heterogeneous LiDAR sensors. Comprehensive evaluations on four large-scale datasets (Oxford RobotCar, NUS Inhouse, HeLiPR, MulRan) demonstrate that RTR-Net not only achieves state-of-the-art results but also stands out as a versatile, plug-and-play module. It is compatible with existing LPR methods—whether region-based or sparse voxelization-based—enhancing their localization accuracy in challenging conditions without requiring structural modifications or retraining. Further experiments on heterogeneous LPR and long-term environmental variations validate RTR-Net’s robustness, achieving leading performance across sensor types and temporal shifts. The proposed regional sampling method effectively alleviates FoV disparities, demonstrating broad applicability within current LPR frameworks.